Lesson 3 — AI, machine learning and deep learning
These three terms are used interchangeably in the press and they are not interchangeable at all. They are nested, like Russian dolls, and knowing which one you are talking about tells you what a project will cost.
The three circles
Artificial Intelligence is the outermost circle: the objective of making machines do intelligent things, by any means. It includes machine learning and also older approaches — the expert systems from lesson 1, search algorithms, automated planning. A chess engine using brute-force search with no learning is still AI.
Machine learning is the dominant subset: the machine derives its rules from examples. This is where roughly every system called AI in 2026 sits.
Deep learning is a branch of machine learning using neural networks with many stacked layers. It dominates whenever the data is raw and unstructured — images, sound, text.
Generative AI is a use of deep learning that produces new content rather than a label. ChatGPT, Claude, Midjourney and Stable Diffusion all live here.
The distinction that actually matters in practice
Beyond taxonomy, one difference decides how a project runs: who invents the features.
Classical machine learning: you describe, it decides
Suppose you want to predict whether a customer will churn. You hand the model a table you built yourself: months of tenure, number of support tickets, average monthly spend, last login date. The model learns to weigh those columns. You chose which columns matter.
This is called feature engineering, and in classical machine learning it is where most of the work and most of the value sit. A domain expert who knows that "three support tickets in one week" is the real warning sign will beat a mediocre expert with a better algorithm every time.
Deep learning: it finds the features itself
Now suppose you want to detect a tumour on a scan. There is no table. There are pixels. Nobody can write down the columns, and asking a radiologist to enumerate the visual features they use runs straight back into Polanyi's paradox from lesson 1.
A deep network builds the intermediate representations itself, layer by layer. Examining a trained image network shows a striking progression: the first layers respond to edges and orientations, the middle layers to textures and simple shapes, the deeper layers to whole objects. No one programmed that hierarchy. It emerged from the data.
That is the real superpower of deep learning, and also the source of its cost.
| Classical machine learning | Deep learning | |
|---|---|---|
| Data type | tables, columns, structured | images, sound, text, raw |
| Features | you design them | the model discovers them |
| Data volume needed | hundreds to thousands of rows | tens of thousands to millions |
| Training hardware | an ordinary laptop | a GPU, sometimes several |
| Training time | seconds to minutes | hours to weeks |
| Explainability | often readable, sometimes fully | very difficult |
| Typical accuracy on tables | often better | rarely better |
The mistake beginners make
Because deep learning gets the headlines, the instinct is to reach for a neural network every time. On tabular data — the spreadsheets that make up most of a company's information — this is usually the wrong call.
On structured tables, gradient-boosted tree ensembles such as XGBoost or LightGBM regularly beat neural networks, train in seconds instead of hours, run on any machine and produce a readable list of which columns drove the decision. This is not folklore; it holds across most published benchmarks on tabular problems.
Raw and unstructured data — pixels, audio waveforms, free text — call for deep learning, because nobody can hand-craft the features. Structured tables call for a tree ensemble first. If it is good enough, and it very often is, you have saved weeks of work and kept a model you can explain to an auditor.
Where language models sit
A large language model is deep learning, built on a specific architecture called the transformer, introduced in 2017. What makes it feel different from earlier AI is not the underlying principle but two facts of scale:
- It was trained on a corpus measured in trillions of words, a substantial slice of publicly available human writing.
- It has hundreds of billions of parameters, enough capacity to retain fine-grained regularities from that corpus.
Its training objective is famously simple: predict the next word. Everything else — answering questions, writing code, translating, summarising — is a side effect of doing that one thing extremely well at enormous scale. The LLM discovery course unpacks this properly.
In three sentences
AI is the goal, machine learning is the dominant method, and deep learning is the branch of machine learning that handles raw data by discovering its own features. The practical dividing line is who invents the features: you do in classical machine learning, the network does in deep learning. Reach for deep learning on images, sound and text, and start with a tree ensemble on tables, where it usually wins anyway.